What is the optimal approach to analyse ventilator-free days? A simulation study

Critical Care volume 29, Article number: 251 (2025)

Background

Ventilator-free days (VFDs) are a composite outcome in critical care research, reflecting both survival and mechanical ventilation duration. However, analysis methods for VFDs are inconsistent, with some focusing on counts and others on time-to-event outcomes, while other approaches such as the multistate model and the win ratio have emerged. We aimed to evaluate various statistical models through simulations to identify the optimal approach for analysing VFDs.

Methods

First, 16 datasets of 300 individuals were simulated, comparing a control group to an intervention with varying survival rates and ventilation durations. Various statistical models were evaluated for statistical power and Type I error rate. Four clinical trial datasets (LIVE study, NCT02149589; ARMA study, NCT00000579; ACURASYS study, NCT00299650; COVIDICUS study, NCT04344730) were then used to apply the same statistical models to analyse VFDs. Twelve statistical methods were evaluated, including count-based, time-to-event approaches, and the win-ratio. Additionally, sensitivity analyses were conducted.

Results

Most statistical methods effectively controlled Type I error rate, except for the zero-inflated and hurdle Poisson/negative binomial count submodels, as well as the cause-specific Cox regression model for death. The power to detect survival benefit and ventilation duration effects varied, with time-to-event approaches, the Mann–Whitney test, the proportional odds model and the win ratio generally performing best. Similar results were observed in sensitivity analyses. In the real datasets, the multistate model, the Mann–Whitney test, the proportional odds model and the win ratio generally showed a significant association between VFDs and randomisation groups.

Conclusions

The multistate model could be recommended as the optimal approach for analysing VFDs, as it outperformed the other methods and offers a more interpretable effect size than the proportional odds model and the win ratio.